COMPUTATIONAL INTELLIGENCE (INTRODUCTION TO MACHINE LEARNING)

Transcription

1 COMPUTATIONAL INTELLIGENCE (INTRODUCTION TO MACHINE LEARNING) SS 18 2 VO UE Institute for Theoretical Computer Science (IGI) TU Graz, Inffeldgasse 16b / first floor Institute for Signal Processing and Speech Communication (SPSC) TU Graz, Inffeldgasse 16c / ground floor

2 Organization Lecture / VO: Tuesday, 11:00, HS i13 Anand Subramoney and Guillaume Bellec (IGI) Assoc. Prof. Dr. Franz Pernkopf (SPSC) Practical / UE: First practical on Friday 9th of March, HS i11 12:30-13:30 if your last name starts with A-L 14:00-15:00 if your last name starts with M-Z Part I Part II Anand Subramoney and Guillaume Bellec (IGI) Part I Dipl.-Ing. Christian Knoll (SPSC) Part II Homework in teams of up to 3 (use newsgroup to form teams) Website: Newsgroup: tu-graz.lv.ci

3 Organization Lecture / VO: Class cancelled on the 13 th of May Practical / UE: Class cancelled on the 16 th of May

4 Organization Office hours: Both Anand and Guillaume: Time: Every Tuesday 14:00 15:00 Place: Our offices at Inffeldgasse 16b/1 Exam: Written exam for this year s course: July onwards Exam has two parts: IGI (first half of semester) + SPSC (second half) Language: English Positive grade if positive on both parts!

5 Materials (for IGI part) No textbook required Lecture slides and further reading on tech center Materials for further study: Online Machine Learning course coursera udacity de.udacity.com/course/intro-to-machine-learning--ud120 Book by C. Bishop, Pattern Recognition and Machine Learning, Springer For SPSC part (second half): Announced by Franz Pernkopf

6 Acknowledgments IGI Slides based on material from Stefan Häusler (IGI), Zeno Jonke (IGI), David Sontag (NYU), Andrew Ng (Stanford), Xiaoli Fern (Oregon State)

7 INTRODUCTION + MOTIVATION

8 Machine Learning Grew out of Artificial Intelligence

9 What is Artificial Intelligence? Source -- Artificial Intelligence: A Modern Approach by Stuart Russell and Peter Norvig

10 But what really is AI? Turing test

11 Turing test AI You ll know it when you see it

12 Components of AI Natural language processing Knowledge representation Automated reasoning Machine learning Computer vision Robotics -- Russel and Norvig

13 Machine Learning Grew out of Artificial Intelligence The ability to adapt to new circumstances and to detect and extrapolate patterns Russel and Norvig Arthur Samuel (1959). Field of study that gives computers the ability to learn without being explicitly programmed.

14 When do we need computers to learn? When human expert knowledge is missing For example, predicting whether some new substance could be an effective treatment for a disease When humans can only do it intuitively Flying a helicopter Recognize visual objects Natural language processing When we need to learn about something that changes frequently Stock market analysis Weather forecasting Computer network routing Customized learning Spam filters, movie/product recommendations

15 Applications of Machine learning Machine learning is used in a wide range of fields including: Bio-informatics Brain-Machine interfaces Computational finance Game playing Information Retrieval Internet fraud detection Medical diagnosis Natural language processing Online advertising Recommender systems Robot locomotion Search engines Sentiment analysis Software engineering Speech and handwriting recognition Stock market analysis Economics and Finance Credit card fraud detection.

16 Autonomous car Waymo/Alphabet + UK, France, Switzerland, Singapore

17 Bipedal robot ATLAS (Boston Dynamics/Alphabet) (three month ago) (last week)

18 AI for robotics OpenAI (2016) robots cannot only learn from accelerated simulated environments

19 Web search

20 Image search Google image search

21 Face recognition Facebook iphoto Cameras, etc. Microsoft cognitive services From face, can recognize age, gender, emotions! e-services/

22 Scene and text recognition Microsoft Seeing AI project

23 Machine Translation Skype and PowerPoint real-time translation (Microsoft)

24 Learning to reason Human level performance at video games from ATARI 2600 (Google Deep mind 2015) Beating world champion of GO (G. Deepmind 2016) Beating champion chest program (G. Deep mind 2017)

25 Brain Computer Interface Neural Dust tiny neural implants from Berkley (2016) (not much AI in BCI for now but it s coming)

26 CLASSICAL PROBLEMS AND APPLICATIONS

27 Recommender systems

28 Spam filtering "Spam in started to become a problem when the Internet was opened up to the general public in the mid-1990s. It grew exponentially over the following years, and today composes some 80 to 85% of all the in the world, by a "conservative estimate". Source: data prediction Spam vs. Not Spam

29 Data visualization (Embedding images) Images have thousands or millions of pixels. Can we give each image a coordinate, such that similar images are near each other? [Saul & Roweis 03]

30 Clustering Clustering data into similar groups

31 Clustering images Set of images

32 Growth of Machine Learning Preferred approach to Speech recognition, Natural language processing Computer vision Robot control Computational biology. Accelerating trend Big data (data mining) Improved algorithms Faster computers Availability of good open-source software and datasets

33 Some of the future challenges The scientific challenges Learning from fewer data (1-shot learning) Generalization Energy efficient hardware and algorithms Understanding animal intelligence Ethical issues of AI Privacy Intelligent weapons Replacing artisans with robots

34 COURSE CONTENT

35 What we will cover IGI part: Introduction Linear regression Non-linear basis functions Logistic regression Under- and over-fitting Model selection k-nn Cross-validation Regularization Neural networks SVM Kernel methods Multiclass classification SPSC part: Parametric & non-parametric density estimation Bayes classifier Gaussian mixture model K-means Markov model & Hidden Markov model Graphical models PCA LDA

36 INTRODUCTION: TYPES OF ML ALGORITHMS

37 Types of Machine Learning algorithms Supervised learning Given: Goal: Training examples with target values Predict target values for new examples Examples: optical character recognition, speech recognition, etc. Unsupervised learning Given: Goal: Training examples without target values Detect and extract structure from data Examples: clustering, segmentation, embedding (visualization), compression, automatic speaker separation Learning from examples (data) Reinforcement learning (not in this course) Given: Feedback (reward/cost) during Goal: trial-and-error episodes Maximize Reward/minimize cost Examples: learning to control a robot/car/helicopter etc. see Master s course Autonomously Learning Systems Learning by doing (trial and error)

38 Supervised Learning: Example Learn to predict output from input (learning from examples) Target values (output) can be continuous (regression) or discrete (classification) E.g. predict the risk level (high vs. low) of a loan applicant based on income and savings Applications: Spam filters Character recognition Speech recognition Collaborative filtering (predicting if a customer will be interested in an advertisement ) Medical diagnosis

39 Unsupervised Learning: Example 90% of collected data is unlabeled Ex. Find patterns and structure in data Clustering art

40 Unsupervised Learning: Applications Market partition: divide a market into distinct subsets of customers Find clusters of similar customers, where each cluster may conceivably be selected as a market target to be reached with a distinct marketing strategy Data representation: Image, document, web clustering Automatic organization of pictures Generate a categorized view of a collection of documents For organizing search results etc. Bioinformatics Clustering the genes based on their expression profile Find clusters of similarly regulated genes functional groups

41 INTRODUCTION: SUPERVISED LEARNING Regression and classification

42 Total income Simple regression example Million USD Top 50 movies (top first weekends) Avengers The Dark Knight X-Men Origins: Wolverine Million USD First weekend income Data source: The Hunger Games: Catching Fire : 158 Mio. USD on opening weekend. How much in total? Predicted: ~418 Mio., Actual: 424 Mio.

43 y Simple regression example (cont d) Data set: Input i, Output First weekend Total Avengers Iron Man x Harry Potter and the Deathly The Dark Knight Rises m data points (data samples) The Dark Knight

44 y Simple regression example (cont d) 700 Data set: Training set x Hypothesis Learning algorithm Test input x Hypothesis h Prediction Parameters

45 Non-linear regression y training data regression x x Non-linear hypothesis, for example Training set Learning algorithm Test input x Hypothesis h Prediction

46 Regression with multiple inputs y y linear hypothesis non-linear hypothesis

47 Multiple inputs continued y Training set i Test input Learning algorithm Hypothesis h Prediction

48 0 1 Simple classification example i labeled data Tumor size (mm) x Malignant? y benign malignant decision boundary (N) (Y) Tumor size (x) (N) (Y) (Y) labels Example hypothesis: 1 if x >

49 Age (x2) Classification with multiple inputs linear decision boundary i Tumor size (mm) Age Malign ant? x1 x2 y (N) (Y) (N) (Y) (Y) Tumor size (x1) benign malignant

50 Age (x2) Age (x2) Non-linear classification linear decision boundary non-linear decision boundary Tumor size (x1) Tumor size (x1) Both hypotheses fit the data quite well. Which one fits the training data better? Which one would you trust more for prediction?

51 Supervised learning (Regr., Class.) Discrete vs. continuous outputs (classification vs. regression) Training set In the next few classes we ll cover: Learning algorithm Learning algorithms for regression and classification (linear regression, neural nets, SVMs, etc.) Test input Hypothesis h Prediction Supervised learning in practice (overfitting, etc.)

52 How to extend to images or sound? Find the best way to represent the data as vectors (i.e. tables of numbers) Light intensity of each pixel for images, time-wise amplitude of air pressure for sounds Knowing the data structure helps to design better representations. When the data is compressed into a lower dimensions recognition is made easier.

53 What is next? Linear regression Gradient descent Non-linear basis functions

54 Supervised, unsupervised or Reinforcement Learning?

55 Regression or classification?